Small target detection based on multimodal data fusion

Object detection is a fundamental task in remote sensing with wide applications in autonomous vehicle navigation, surveillance, and search-and-rescue operations. Remote sensing images, typically captured by satellites, aircraft, and UAV sensors, are of great significance in both military and civilian domains due to their extensive coverage and ease of collection. However, several challenges arise in the capture of remote sensing images: first, objects are often very small relative to the size of the image, leading to limited feature information; second, objects can be oriented arbitrarily, requiring detection algorithms to learn rotation-invariant features for better performance; and third, noise and occlusions need to be handled by detection algorithms. In this paper, we propose MR-YOLOV5, a convolutional network specifically designed for real-time object detection in remote sensing images. MR-YOLOV5 is capable of detecting small objects with arbitrary orientations and tight spacing, as well as performing large-scale object detection. Additionally, by integrating multi-modal remote sensing data, the network’s performance in detecting challenging objects is enhanced. Experimental results on the publicly available VEDAI dataset demonstrate the effectiveness of the proposed algorithm.